Google releases Gemma open models

Background

An open model, in this context, is one whose trained weights are published so that people outside the organisation that built it can download, run and adapt them. Google said Gemma was built from the same research and technology as Gemini.^[1]

What was released

On 21 February 2024 Google released Gemma, a family of open models.^[1] The family came in two sizes, 2B and 7B, and each size was offered both as a pre-trained version and as an instruction-tuned version.^[1]

The numbers in the names are parameter counts: roughly two billion parameters for 2B and seven billion for 7B. Parameters are the learned values inside a network; more of them generally mean a larger model that costs more memory and compute to run. The two versions differ in what they are for. A pre-trained model has learned from data but is not specifically tuned to follow requests; an instruction-tuned model has been trained further on examples of instructions and the responses expected for them, so it answers a question directly instead of merely continuing text.

How it was distributed

Weights were distributed through Kaggle and Hugging Face.^[1] Alongside them, Google provided ready-to-use Colab and Kaggle notebooks, plus integration with tools such as Hugging Face.^[1] Google also offered a free tier for Colab notebooks to run Gemma.^[1]

Taken together, these choices describe a particular route to the weights: not a single download page, but two established model-hosting platforms, notebook environments that open in a browser, and connectors to tools people already used. A notebook is a document that mixes runnable code with text, so a first run can happen without assembling an environment by hand.

Where it can run

Google said Gemma can run on a laptop, a workstation, or Google Cloud.^[1] The smallest size in the family, 2B, is the one that makes a laptop plausible at all: a few billion parameters fit into the memory of ordinary hardware in a way that much larger models do not. The same weights can also be moved to a workstation or to cloud instances when more compute is needed.

The terms of use

The terms of use allowed responsible commercial usage and distribution by all organisations, regardless of size.^[1] That wording matters for who can build on the models. A licence that is limited by company size, or that excludes commercial work, forces small teams and individual developers to negotiate separately or stay away; terms that apply to any organisation, with responsibility as the condition, let the same artefact be used by a student, a startup and a large company without a bespoke agreement.

Why the release mattered

Gemma arrived as an open family released alongside a company's closed model line, and the details of the release are what shaped its use. The 2B size and the stated ability to run on a laptop lowered the hardware barrier.^[1] The commercial terms removed an organisational barrier.^[1] Distribution through Kaggle and Hugging Face put the weights where people already looked for models, and the ready-made notebooks and integrations shortened the distance between downloading and running something.^[1] The free Colab tier made the first experiment cost nothing.^[1]

For learners, that combination is the practical meaning of an open model: the thing can be inspected and run on hardware you already have, under terms that do not depend on who you are.